Multi image noise estimation and denoising

Photon accumulation on a fixed surface is the essence of photography. In the times of chemical photography this accumulation required the camera to move as little as possible, and the scene to be still. Yet, most recent reflex and compact cameras propose a burst mode, permitting to capture quickly dozens of short exposure images of a scene instead of a single one. This new feature permits in principle to obtain by simple accumulation high quality photographs in dim light, with no motion or aperture blur. It also gives the right data for an accurate noise model. Yet, both goals are attainable only if an accurate cross-registration of the burst images has been performed. The difficulty comes from the non negligible image deformations caused by the slightest camera motion, in front of a 3D scene, and from the light variations or motions in the scene. This paper proposes a numerical processing chain permitting to achieve jointly the two mentioned goals: an accurate noise model for the camera, which is used crucially to obtain a state of the art multi-images denoising. The key feature of the proposed processing chain is a reliable multi-image noise estimator, whose accuracy will be demonstrated by three different procedures. Thanks to the signal dependent noise model obtained from the burst itself, a faithful detection of the well registered pixels can be made. The denoising by simple accumulation of these pixels, which are an overwhelming majority, permits to extend the Nicephore Niepce photon accumulation method to image bursts. The denoising performance by accumulation is shown to reach the theoretical limit, namely a square root of n denoising factor for n frames. Comparison with state of the art denoising algorithms will be shown on several bursts taken with reflex cameras in dim light.